Hyperspectral estimation of magnesium content in Yunyan 87 and Zhongyan 100 tobacco leaves using machine learning

Hyperspectral remote sensing provides a rapid and non-destructive approach for monitoring plant nutrient status; however, its application for magnesium (Mg) estimation in flue-cured tobacco remains limited. In this study, two cultivars, Yunyan 87 and Zhongyan 100, were grown in a hydroponic system with five Mg concentration gradients (0, 0.2, 1, 5, and 25 mmol L^-1). Hyperspectral reflectance data of fresh leaves were collected at different growth stages. Three preprocessing methods, including first derivative (FD), standard normal variate (SNV), and multiplicative scatter correction (MSC), were applied, and partial least squares regression (PLSR) was used to identify the optimal preprocessing strategy. Characteristic wavelengths were selected using competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), and genetic algorithm (GA), and were further combined with extreme learning machine (ELM), support vector regression (SVR), and radial basis function (RBF) neural network models to estimate Mg content. The results showed that spectral preprocessing significantly improved the relationship between hyperspectral data and Mg content, with optimal methods varying across cultivars and growth stages. Selected wavelengths were mainly located in the near-infrared region. The developed models achieved high prediction accuracy, particularly during the middle and late growth stages, where the coefficients of determination (R²) of all test sets exceeded 0.90. In addition, Yunyan 87 exhibited higher prediction accuracy than Zhongyan 100. These findings demonstrate that hyperspectral technology combined with feature wavelength selection and machine learning enables accurate and non-destructive estimation of Mg content in flue-cured tobacco leaves, providing a reliable tool for Mg nutrition diagnosis and precision management. However, further validation under diverse field conditions is required to enhance model robustness.